基于ICF和多级分类器的交通标志检测  

Traffic signs detection based on ICF and multi-class classifier

作　　者：郑天宇[1] 王年[1] 唐俊[1] 

机构地区：[1]安徽大学电子与信息工程学院,安徽合肥230601

出　　处：《传感器与微系统》2017年第6期134-137,共4页Transducer and Microsystem Technologies

基　　金：国家自然科学基金资助项目(61172127);安徽省自然科学基金资助项目(1508085MF120)

摘　　要：针对传统的基于积分通道特征(ICF)和Adaboost交通标志检测算法,召回率过低和误检率过高的问题,提出了一种两阶段交通标志检测方法。第一阶段对ICF进行谱聚类并结合Adaboost算法学习得到目标感兴趣区域(ROI);第二阶段对所获得的感兴趣区域进行直方图均衡化,利用尺度不变特征变换(SIFT)描述子与支持向量机(SVM)分类器相结合,提高了目标区域检测的准确性。通过德国交通标志数据集(GTSDB)的验证,结果表明:采用SICF-Adaboost+SIFT-SVM构建的交通标志级联分类器检测算法相对于传统的ICF-Adaboost算法召回率高且误检率低,适用于真实场景下的交通标志检测。Aiming at problem of low recall rate and high error rate of traditional traffic sign detection algorithm based on integral channel feature（ ICF） and Adaboost,propose a two-stage traffic sign detection method. In the first stage,using spectral clustered integral channel features obtained by spectral clustering are combined with Adaboost to learn the overall detection model,which is applied to the input image to obtain region of interest（ ROI）. In the second stage,histogram equalization is imposed on ROI,and then a shape classifier using support vector machine（ SVM） is employed to filter candidate object regions obtained in the former stage to remove the false positives. Experimental results show that the proposed SICF-Adaboost ＋ SIFT-SVM detection algorithm method built upon a cascade classifier framework possesses higher detection rate compared with the traditional ICF Adaboost detect algorithm in dealing with high light intensity,motion blur,fog,and noisy similar object,which is suitable for traffic sign detection in real-world scenes.

关 键 词：交通标志检测 形状分类器 谱聚类 积分通道特征 感兴趣区域 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]